OpenAI just shipped a model that writes exploits and they called it alignment
GPT-6 Cyber can find zero-days and write working shell code. The alignment strategy is hoping it refuses when bad actors ask.
Read the note 1 min read
From the notebook
2 notes on this topic.
GPT-6 Cyber can find zero-days and write working shell code. The alignment strategy is hoping it refuses when bad actors ask.
OpenAI open-sourced the internal security guidelines they used when building Codex. Turns out threat modelling an AI code generator is different from threat modelling a database.